Black Lives Matter. Support the Equal Justice Initiative.

Source file src/runtime/malloc.go

Documentation: runtime

     1  // Copyright 2014 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Memory allocator.
     6  //
     7  // This was originally based on tcmalloc, but has diverged quite a bit.
     8  // http://goog-perftools.sourceforge.net/doc/tcmalloc.html
     9  
    10  // The main allocator works in runs of pages.
    11  // Small allocation sizes (up to and including 32 kB) are
    12  // rounded to one of about 70 size classes, each of which
    13  // has its own free set of objects of exactly that size.
    14  // Any free page of memory can be split into a set of objects
    15  // of one size class, which are then managed using a free bitmap.
    16  //
    17  // The allocator's data structures are:
    18  //
    19  //	fixalloc: a free-list allocator for fixed-size off-heap objects,
    20  //		used to manage storage used by the allocator.
    21  //	mheap: the malloc heap, managed at page (8192-byte) granularity.
    22  //	mspan: a run of in-use pages managed by the mheap.
    23  //	mcentral: collects all spans of a given size class.
    24  //	mcache: a per-P cache of mspans with free space.
    25  //	mstats: allocation statistics.
    26  //
    27  // Allocating a small object proceeds up a hierarchy of caches:
    28  //
    29  //	1. Round the size up to one of the small size classes
    30  //	   and look in the corresponding mspan in this P's mcache.
    31  //	   Scan the mspan's free bitmap to find a free slot.
    32  //	   If there is a free slot, allocate it.
    33  //	   This can all be done without acquiring a lock.
    34  //
    35  //	2. If the mspan has no free slots, obtain a new mspan
    36  //	   from the mcentral's list of mspans of the required size
    37  //	   class that have free space.
    38  //	   Obtaining a whole span amortizes the cost of locking
    39  //	   the mcentral.
    40  //
    41  //	3. If the mcentral's mspan list is empty, obtain a run
    42  //	   of pages from the mheap to use for the mspan.
    43  //
    44  //	4. If the mheap is empty or has no page runs large enough,
    45  //	   allocate a new group of pages (at least 1MB) from the
    46  //	   operating system. Allocating a large run of pages
    47  //	   amortizes the cost of talking to the operating system.
    48  //
    49  // Sweeping an mspan and freeing objects on it proceeds up a similar
    50  // hierarchy:
    51  //
    52  //	1. If the mspan is being swept in response to allocation, it
    53  //	   is returned to the mcache to satisfy the allocation.
    54  //
    55  //	2. Otherwise, if the mspan still has allocated objects in it,
    56  //	   it is placed on the mcentral free list for the mspan's size
    57  //	   class.
    58  //
    59  //	3. Otherwise, if all objects in the mspan are free, the mspan's
    60  //	   pages are returned to the mheap and the mspan is now dead.
    61  //
    62  // Allocating and freeing a large object uses the mheap
    63  // directly, bypassing the mcache and mcentral.
    64  //
    65  // If mspan.needzero is false, then free object slots in the mspan are
    66  // already zeroed. Otherwise if needzero is true, objects are zeroed as
    67  // they are allocated. There are various benefits to delaying zeroing
    68  // this way:
    69  //
    70  //	1. Stack frame allocation can avoid zeroing altogether.
    71  //
    72  //	2. It exhibits better temporal locality, since the program is
    73  //	   probably about to write to the memory.
    74  //
    75  //	3. We don't zero pages that never get reused.
    76  
    77  // Virtual memory layout
    78  //
    79  // The heap consists of a set of arenas, which are 64MB on 64-bit and
    80  // 4MB on 32-bit (heapArenaBytes). Each arena's start address is also
    81  // aligned to the arena size.
    82  //
    83  // Each arena has an associated heapArena object that stores the
    84  // metadata for that arena: the heap bitmap for all words in the arena
    85  // and the span map for all pages in the arena. heapArena objects are
    86  // themselves allocated off-heap.
    87  //
    88  // Since arenas are aligned, the address space can be viewed as a
    89  // series of arena frames. The arena map (mheap_.arenas) maps from
    90  // arena frame number to *heapArena, or nil for parts of the address
    91  // space not backed by the Go heap. The arena map is structured as a
    92  // two-level array consisting of a "L1" arena map and many "L2" arena
    93  // maps; however, since arenas are large, on many architectures, the
    94  // arena map consists of a single, large L2 map.
    95  //
    96  // The arena map covers the entire possible address space, allowing
    97  // the Go heap to use any part of the address space. The allocator
    98  // attempts to keep arenas contiguous so that large spans (and hence
    99  // large objects) can cross arenas.
   100  
   101  package runtime
   102  
   103  import (
   104  	"runtime/internal/atomic"
   105  	"runtime/internal/math"
   106  	"runtime/internal/sys"
   107  	"unsafe"
   108  )
   109  
   110  const (
   111  	debugMalloc = false
   112  
   113  	maxTinySize   = _TinySize
   114  	tinySizeClass = _TinySizeClass
   115  	maxSmallSize  = _MaxSmallSize
   116  
   117  	pageShift = _PageShift
   118  	pageSize  = _PageSize
   119  	pageMask  = _PageMask
   120  	// By construction, single page spans of the smallest object class
   121  	// have the most objects per span.
   122  	maxObjsPerSpan = pageSize / 8
   123  
   124  	concurrentSweep = _ConcurrentSweep
   125  
   126  	_PageSize = 1 << _PageShift
   127  	_PageMask = _PageSize - 1
   128  
   129  	// _64bit = 1 on 64-bit systems, 0 on 32-bit systems
   130  	_64bit = 1 << (^uintptr(0) >> 63) / 2
   131  
   132  	// Tiny allocator parameters, see "Tiny allocator" comment in malloc.go.
   133  	_TinySize      = 16
   134  	_TinySizeClass = int8(2)
   135  
   136  	_FixAllocChunk = 16 << 10 // Chunk size for FixAlloc
   137  
   138  	// Per-P, per order stack segment cache size.
   139  	_StackCacheSize = 32 * 1024
   140  
   141  	// Number of orders that get caching. Order 0 is FixedStack
   142  	// and each successive order is twice as large.
   143  	// We want to cache 2KB, 4KB, 8KB, and 16KB stacks. Larger stacks
   144  	// will be allocated directly.
   145  	// Since FixedStack is different on different systems, we
   146  	// must vary NumStackOrders to keep the same maximum cached size.
   147  	//   OS               | FixedStack | NumStackOrders
   148  	//   -----------------+------------+---------------
   149  	//   linux/darwin/bsd | 2KB        | 4
   150  	//   windows/32       | 4KB        | 3
   151  	//   windows/64       | 8KB        | 2
   152  	//   plan9            | 4KB        | 3
   153  	_NumStackOrders = 4 - sys.PtrSize/4*sys.GoosWindows - 1*sys.GoosPlan9
   154  
   155  	// heapAddrBits is the number of bits in a heap address. On
   156  	// amd64, addresses are sign-extended beyond heapAddrBits. On
   157  	// other arches, they are zero-extended.
   158  	//
   159  	// On most 64-bit platforms, we limit this to 48 bits based on a
   160  	// combination of hardware and OS limitations.
   161  	//
   162  	// amd64 hardware limits addresses to 48 bits, sign-extended
   163  	// to 64 bits. Addresses where the top 16 bits are not either
   164  	// all 0 or all 1 are "non-canonical" and invalid. Because of
   165  	// these "negative" addresses, we offset addresses by 1<<47
   166  	// (arenaBaseOffset) on amd64 before computing indexes into
   167  	// the heap arenas index. In 2017, amd64 hardware added
   168  	// support for 57 bit addresses; however, currently only Linux
   169  	// supports this extension and the kernel will never choose an
   170  	// address above 1<<47 unless mmap is called with a hint
   171  	// address above 1<<47 (which we never do).
   172  	//
   173  	// arm64 hardware (as of ARMv8) limits user addresses to 48
   174  	// bits, in the range [0, 1<<48).
   175  	//
   176  	// ppc64, mips64, and s390x support arbitrary 64 bit addresses
   177  	// in hardware. On Linux, Go leans on stricter OS limits. Based
   178  	// on Linux's processor.h, the user address space is limited as
   179  	// follows on 64-bit architectures:
   180  	//
   181  	// Architecture  Name              Maximum Value (exclusive)
   182  	// ---------------------------------------------------------------------
   183  	// amd64         TASK_SIZE_MAX     0x007ffffffff000 (47 bit addresses)
   184  	// arm64         TASK_SIZE_64      0x01000000000000 (48 bit addresses)
   185  	// ppc64{,le}    TASK_SIZE_USER64  0x00400000000000 (46 bit addresses)
   186  	// mips64{,le}   TASK_SIZE64       0x00010000000000 (40 bit addresses)
   187  	// s390x         TASK_SIZE         1<<64 (64 bit addresses)
   188  	//
   189  	// These limits may increase over time, but are currently at
   190  	// most 48 bits except on s390x. On all architectures, Linux
   191  	// starts placing mmap'd regions at addresses that are
   192  	// significantly below 48 bits, so even if it's possible to
   193  	// exceed Go's 48 bit limit, it's extremely unlikely in
   194  	// practice.
   195  	//
   196  	// On 32-bit platforms, we accept the full 32-bit address
   197  	// space because doing so is cheap.
   198  	// mips32 only has access to the low 2GB of virtual memory, so
   199  	// we further limit it to 31 bits.
   200  	//
   201  	// On ios/arm64, although 64-bit pointers are presumably
   202  	// available, pointers are truncated to 33 bits. Furthermore,
   203  	// only the top 4 GiB of the address space are actually available
   204  	// to the application, but we allow the whole 33 bits anyway for
   205  	// simplicity.
   206  	// TODO(mknyszek): Consider limiting it to 32 bits and using
   207  	// arenaBaseOffset to offset into the top 4 GiB.
   208  	//
   209  	// WebAssembly currently has a limit of 4GB linear memory.
   210  	heapAddrBits = (_64bit*(1-sys.GoarchWasm)*(1-sys.GoosIos*sys.GoarchArm64))*48 + (1-_64bit+sys.GoarchWasm)*(32-(sys.GoarchMips+sys.GoarchMipsle)) + 33*sys.GoosIos*sys.GoarchArm64
   211  
   212  	// maxAlloc is the maximum size of an allocation. On 64-bit,
   213  	// it's theoretically possible to allocate 1<<heapAddrBits bytes. On
   214  	// 32-bit, however, this is one less than 1<<32 because the
   215  	// number of bytes in the address space doesn't actually fit
   216  	// in a uintptr.
   217  	maxAlloc = (1 << heapAddrBits) - (1-_64bit)*1
   218  
   219  	// The number of bits in a heap address, the size of heap
   220  	// arenas, and the L1 and L2 arena map sizes are related by
   221  	//
   222  	//   (1 << addr bits) = arena size * L1 entries * L2 entries
   223  	//
   224  	// Currently, we balance these as follows:
   225  	//
   226  	//       Platform  Addr bits  Arena size  L1 entries   L2 entries
   227  	// --------------  ---------  ----------  ----------  -----------
   228  	//       */64-bit         48        64MB           1    4M (32MB)
   229  	// windows/64-bit         48         4MB          64    1M  (8MB)
   230  	//      ios/arm64         33         4MB           1  2048  (8KB)
   231  	//       */32-bit         32         4MB           1  1024  (4KB)
   232  	//     */mips(le)         31         4MB           1   512  (2KB)
   233  
   234  	// heapArenaBytes is the size of a heap arena. The heap
   235  	// consists of mappings of size heapArenaBytes, aligned to
   236  	// heapArenaBytes. The initial heap mapping is one arena.
   237  	//
   238  	// This is currently 64MB on 64-bit non-Windows and 4MB on
   239  	// 32-bit and on Windows. We use smaller arenas on Windows
   240  	// because all committed memory is charged to the process,
   241  	// even if it's not touched. Hence, for processes with small
   242  	// heaps, the mapped arena space needs to be commensurate.
   243  	// This is particularly important with the race detector,
   244  	// since it significantly amplifies the cost of committed
   245  	// memory.
   246  	heapArenaBytes = 1 << logHeapArenaBytes
   247  
   248  	// logHeapArenaBytes is log_2 of heapArenaBytes. For clarity,
   249  	// prefer using heapArenaBytes where possible (we need the
   250  	// constant to compute some other constants).
   251  	logHeapArenaBytes = (6+20)*(_64bit*(1-sys.GoosWindows)*(1-sys.GoarchWasm)*(1-sys.GoosIos*sys.GoarchArm64)) + (2+20)*(_64bit*sys.GoosWindows) + (2+20)*(1-_64bit) + (2+20)*sys.GoarchWasm + (2+20)*sys.GoosIos*sys.GoarchArm64
   252  
   253  	// heapArenaBitmapBytes is the size of each heap arena's bitmap.
   254  	heapArenaBitmapBytes = heapArenaBytes / (sys.PtrSize * 8 / 2)
   255  
   256  	pagesPerArena = heapArenaBytes / pageSize
   257  
   258  	// arenaL1Bits is the number of bits of the arena number
   259  	// covered by the first level arena map.
   260  	//
   261  	// This number should be small, since the first level arena
   262  	// map requires PtrSize*(1<<arenaL1Bits) of space in the
   263  	// binary's BSS. It can be zero, in which case the first level
   264  	// index is effectively unused. There is a performance benefit
   265  	// to this, since the generated code can be more efficient,
   266  	// but comes at the cost of having a large L2 mapping.
   267  	//
   268  	// We use the L1 map on 64-bit Windows because the arena size
   269  	// is small, but the address space is still 48 bits, and
   270  	// there's a high cost to having a large L2.
   271  	arenaL1Bits = 6 * (_64bit * sys.GoosWindows)
   272  
   273  	// arenaL2Bits is the number of bits of the arena number
   274  	// covered by the second level arena index.
   275  	//
   276  	// The size of each arena map allocation is proportional to
   277  	// 1<<arenaL2Bits, so it's important that this not be too
   278  	// large. 48 bits leads to 32MB arena index allocations, which
   279  	// is about the practical threshold.
   280  	arenaL2Bits = heapAddrBits - logHeapArenaBytes - arenaL1Bits
   281  
   282  	// arenaL1Shift is the number of bits to shift an arena frame
   283  	// number by to compute an index into the first level arena map.
   284  	arenaL1Shift = arenaL2Bits
   285  
   286  	// arenaBits is the total bits in a combined arena map index.
   287  	// This is split between the index into the L1 arena map and
   288  	// the L2 arena map.
   289  	arenaBits = arenaL1Bits + arenaL2Bits
   290  
   291  	// arenaBaseOffset is the pointer value that corresponds to
   292  	// index 0 in the heap arena map.
   293  	//
   294  	// On amd64, the address space is 48 bits, sign extended to 64
   295  	// bits. This offset lets us handle "negative" addresses (or
   296  	// high addresses if viewed as unsigned).
   297  	//
   298  	// On aix/ppc64, this offset allows to keep the heapAddrBits to
   299  	// 48. Otherwise, it would be 60 in order to handle mmap addresses
   300  	// (in range 0x0a00000000000000 - 0x0afffffffffffff). But in this
   301  	// case, the memory reserved in (s *pageAlloc).init for chunks
   302  	// is causing important slowdowns.
   303  	//
   304  	// On other platforms, the user address space is contiguous
   305  	// and starts at 0, so no offset is necessary.
   306  	arenaBaseOffset = 0xffff800000000000*sys.GoarchAmd64 + 0x0a00000000000000*sys.GoosAix
   307  	// A typed version of this constant that will make it into DWARF (for viewcore).
   308  	arenaBaseOffsetUintptr = uintptr(arenaBaseOffset)
   309  
   310  	// Max number of threads to run garbage collection.
   311  	// 2, 3, and 4 are all plausible maximums depending
   312  	// on the hardware details of the machine. The garbage
   313  	// collector scales well to 32 cpus.
   314  	_MaxGcproc = 32
   315  
   316  	// minLegalPointer is the smallest possible legal pointer.
   317  	// This is the smallest possible architectural page size,
   318  	// since we assume that the first page is never mapped.
   319  	//
   320  	// This should agree with minZeroPage in the compiler.
   321  	minLegalPointer uintptr = 4096
   322  )
   323  
   324  // physPageSize is the size in bytes of the OS's physical pages.
   325  // Mapping and unmapping operations must be done at multiples of
   326  // physPageSize.
   327  //
   328  // This must be set by the OS init code (typically in osinit) before
   329  // mallocinit.
   330  var physPageSize uintptr
   331  
   332  // physHugePageSize is the size in bytes of the OS's default physical huge
   333  // page size whose allocation is opaque to the application. It is assumed
   334  // and verified to be a power of two.
   335  //
   336  // If set, this must be set by the OS init code (typically in osinit) before
   337  // mallocinit. However, setting it at all is optional, and leaving the default
   338  // value is always safe (though potentially less efficient).
   339  //
   340  // Since physHugePageSize is always assumed to be a power of two,
   341  // physHugePageShift is defined as physHugePageSize == 1 << physHugePageShift.
   342  // The purpose of physHugePageShift is to avoid doing divisions in
   343  // performance critical functions.
   344  var (
   345  	physHugePageSize  uintptr
   346  	physHugePageShift uint
   347  )
   348  
   349  // OS memory management abstraction layer
   350  //
   351  // Regions of the address space managed by the runtime may be in one of four
   352  // states at any given time:
   353  // 1) None - Unreserved and unmapped, the default state of any region.
   354  // 2) Reserved - Owned by the runtime, but accessing it would cause a fault.
   355  //               Does not count against the process' memory footprint.
   356  // 3) Prepared - Reserved, intended not to be backed by physical memory (though
   357  //               an OS may implement this lazily). Can transition efficiently to
   358  //               Ready. Accessing memory in such a region is undefined (may
   359  //               fault, may give back unexpected zeroes, etc.).
   360  // 4) Ready - may be accessed safely.
   361  //
   362  // This set of states is more than is strictly necessary to support all the
   363  // currently supported platforms. One could get by with just None, Reserved, and
   364  // Ready. However, the Prepared state gives us flexibility for performance
   365  // purposes. For example, on POSIX-y operating systems, Reserved is usually a
   366  // private anonymous mmap'd region with PROT_NONE set, and to transition
   367  // to Ready would require setting PROT_READ|PROT_WRITE. However the
   368  // underspecification of Prepared lets us use just MADV_FREE to transition from
   369  // Ready to Prepared. Thus with the Prepared state we can set the permission
   370  // bits just once early on, we can efficiently tell the OS that it's free to
   371  // take pages away from us when we don't strictly need them.
   372  //
   373  // For each OS there is a common set of helpers defined that transition
   374  // memory regions between these states. The helpers are as follows:
   375  //
   376  // sysAlloc transitions an OS-chosen region of memory from None to Ready.
   377  // More specifically, it obtains a large chunk of zeroed memory from the
   378  // operating system, typically on the order of a hundred kilobytes
   379  // or a megabyte. This memory is always immediately available for use.
   380  //
   381  // sysFree transitions a memory region from any state to None. Therefore, it
   382  // returns memory unconditionally. It is used if an out-of-memory error has been
   383  // detected midway through an allocation or to carve out an aligned section of
   384  // the address space. It is okay if sysFree is a no-op only if sysReserve always
   385  // returns a memory region aligned to the heap allocator's alignment
   386  // restrictions.
   387  //
   388  // sysReserve transitions a memory region from None to Reserved. It reserves
   389  // address space in such a way that it would cause a fatal fault upon access
   390  // (either via permissions or not committing the memory). Such a reservation is
   391  // thus never backed by physical memory.
   392  // If the pointer passed to it is non-nil, the caller wants the
   393  // reservation there, but sysReserve can still choose another
   394  // location if that one is unavailable.
   395  // NOTE: sysReserve returns OS-aligned memory, but the heap allocator
   396  // may use larger alignment, so the caller must be careful to realign the
   397  // memory obtained by sysReserve.
   398  //
   399  // sysMap transitions a memory region from Reserved to Prepared. It ensures the
   400  // memory region can be efficiently transitioned to Ready.
   401  //
   402  // sysUsed transitions a memory region from Prepared to Ready. It notifies the
   403  // operating system that the memory region is needed and ensures that the region
   404  // may be safely accessed. This is typically a no-op on systems that don't have
   405  // an explicit commit step and hard over-commit limits, but is critical on
   406  // Windows, for example.
   407  //
   408  // sysUnused transitions a memory region from Ready to Prepared. It notifies the
   409  // operating system that the physical pages backing this memory region are no
   410  // longer needed and can be reused for other purposes. The contents of a
   411  // sysUnused memory region are considered forfeit and the region must not be
   412  // accessed again until sysUsed is called.
   413  //
   414  // sysFault transitions a memory region from Ready or Prepared to Reserved. It
   415  // marks a region such that it will always fault if accessed. Used only for
   416  // debugging the runtime.
   417  
   418  func mallocinit() {
   419  	if class_to_size[_TinySizeClass] != _TinySize {
   420  		throw("bad TinySizeClass")
   421  	}
   422  
   423  	testdefersizes()
   424  
   425  	if heapArenaBitmapBytes&(heapArenaBitmapBytes-1) != 0 {
   426  		// heapBits expects modular arithmetic on bitmap
   427  		// addresses to work.
   428  		throw("heapArenaBitmapBytes not a power of 2")
   429  	}
   430  
   431  	// Copy class sizes out for statistics table.
   432  	for i := range class_to_size {
   433  		memstats.by_size[i].size = uint32(class_to_size[i])
   434  	}
   435  
   436  	// Check physPageSize.
   437  	if physPageSize == 0 {
   438  		// The OS init code failed to fetch the physical page size.
   439  		throw("failed to get system page size")
   440  	}
   441  	if physPageSize > maxPhysPageSize {
   442  		print("system page size (", physPageSize, ") is larger than maximum page size (", maxPhysPageSize, ")\n")
   443  		throw("bad system page size")
   444  	}
   445  	if physPageSize < minPhysPageSize {
   446  		print("system page size (", physPageSize, ") is smaller than minimum page size (", minPhysPageSize, ")\n")
   447  		throw("bad system page size")
   448  	}
   449  	if physPageSize&(physPageSize-1) != 0 {
   450  		print("system page size (", physPageSize, ") must be a power of 2\n")
   451  		throw("bad system page size")
   452  	}
   453  	if physHugePageSize&(physHugePageSize-1) != 0 {
   454  		print("system huge page size (", physHugePageSize, ") must be a power of 2\n")
   455  		throw("bad system huge page size")
   456  	}
   457  	if physHugePageSize > maxPhysHugePageSize {
   458  		// physHugePageSize is greater than the maximum supported huge page size.
   459  		// Don't throw here, like in the other cases, since a system configured
   460  		// in this way isn't wrong, we just don't have the code to support them.
   461  		// Instead, silently set the huge page size to zero.
   462  		physHugePageSize = 0
   463  	}
   464  	if physHugePageSize != 0 {
   465  		// Since physHugePageSize is a power of 2, it suffices to increase
   466  		// physHugePageShift until 1<<physHugePageShift == physHugePageSize.
   467  		for 1<<physHugePageShift != physHugePageSize {
   468  			physHugePageShift++
   469  		}
   470  	}
   471  	if pagesPerArena%pagesPerSpanRoot != 0 {
   472  		print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerSpanRoot (", pagesPerSpanRoot, ")\n")
   473  		throw("bad pagesPerSpanRoot")
   474  	}
   475  	if pagesPerArena%pagesPerReclaimerChunk != 0 {
   476  		print("pagesPerArena (", pagesPerArena, ") is not divisible by pagesPerReclaimerChunk (", pagesPerReclaimerChunk, ")\n")
   477  		throw("bad pagesPerReclaimerChunk")
   478  	}
   479  
   480  	// Initialize the heap.
   481  	mheap_.init()
   482  	mcache0 = allocmcache()
   483  	lockInit(&gcBitsArenas.lock, lockRankGcBitsArenas)
   484  	lockInit(&proflock, lockRankProf)
   485  	lockInit(&globalAlloc.mutex, lockRankGlobalAlloc)
   486  
   487  	// Create initial arena growth hints.
   488  	if sys.PtrSize == 8 {
   489  		// On a 64-bit machine, we pick the following hints
   490  		// because:
   491  		//
   492  		// 1. Starting from the middle of the address space
   493  		// makes it easier to grow out a contiguous range
   494  		// without running in to some other mapping.
   495  		//
   496  		// 2. This makes Go heap addresses more easily
   497  		// recognizable when debugging.
   498  		//
   499  		// 3. Stack scanning in gccgo is still conservative,
   500  		// so it's important that addresses be distinguishable
   501  		// from other data.
   502  		//
   503  		// Starting at 0x00c0 means that the valid memory addresses
   504  		// will begin 0x00c0, 0x00c1, ...
   505  		// In little-endian, that's c0 00, c1 00, ... None of those are valid
   506  		// UTF-8 sequences, and they are otherwise as far away from
   507  		// ff (likely a common byte) as possible. If that fails, we try other 0xXXc0
   508  		// addresses. An earlier attempt to use 0x11f8 caused out of memory errors
   509  		// on OS X during thread allocations.  0x00c0 causes conflicts with
   510  		// AddressSanitizer which reserves all memory up to 0x0100.
   511  		// These choices reduce the odds of a conservative garbage collector
   512  		// not collecting memory because some non-pointer block of memory
   513  		// had a bit pattern that matched a memory address.
   514  		//
   515  		// However, on arm64, we ignore all this advice above and slam the
   516  		// allocation at 0x40 << 32 because when using 4k pages with 3-level
   517  		// translation buffers, the user address space is limited to 39 bits
   518  		// On ios/arm64, the address space is even smaller.
   519  		//
   520  		// On AIX, mmaps starts at 0x0A00000000000000 for 64-bit.
   521  		// processes.
   522  		for i := 0x7f; i >= 0; i-- {
   523  			var p uintptr
   524  			switch {
   525  			case raceenabled:
   526  				// The TSAN runtime requires the heap
   527  				// to be in the range [0x00c000000000,
   528  				// 0x00e000000000).
   529  				p = uintptr(i)<<32 | uintptrMask&(0x00c0<<32)
   530  				if p >= uintptrMask&0x00e000000000 {
   531  					continue
   532  				}
   533  			case GOARCH == "arm64" && GOOS == "ios":
   534  				p = uintptr(i)<<40 | uintptrMask&(0x0013<<28)
   535  			case GOARCH == "arm64":
   536  				p = uintptr(i)<<40 | uintptrMask&(0x0040<<32)
   537  			case GOOS == "aix":
   538  				if i == 0 {
   539  					// We don't use addresses directly after 0x0A00000000000000
   540  					// to avoid collisions with others mmaps done by non-go programs.
   541  					continue
   542  				}
   543  				p = uintptr(i)<<40 | uintptrMask&(0xa0<<52)
   544  			default:
   545  				p = uintptr(i)<<40 | uintptrMask&(0x00c0<<32)
   546  			}
   547  			hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc())
   548  			hint.addr = p
   549  			hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
   550  		}
   551  	} else {
   552  		// On a 32-bit machine, we're much more concerned
   553  		// about keeping the usable heap contiguous.
   554  		// Hence:
   555  		//
   556  		// 1. We reserve space for all heapArenas up front so
   557  		// they don't get interleaved with the heap. They're
   558  		// ~258MB, so this isn't too bad. (We could reserve a
   559  		// smaller amount of space up front if this is a
   560  		// problem.)
   561  		//
   562  		// 2. We hint the heap to start right above the end of
   563  		// the binary so we have the best chance of keeping it
   564  		// contiguous.
   565  		//
   566  		// 3. We try to stake out a reasonably large initial
   567  		// heap reservation.
   568  
   569  		const arenaMetaSize = (1 << arenaBits) * unsafe.Sizeof(heapArena{})
   570  		meta := uintptr(sysReserve(nil, arenaMetaSize))
   571  		if meta != 0 {
   572  			mheap_.heapArenaAlloc.init(meta, arenaMetaSize, true)
   573  		}
   574  
   575  		// We want to start the arena low, but if we're linked
   576  		// against C code, it's possible global constructors
   577  		// have called malloc and adjusted the process' brk.
   578  		// Query the brk so we can avoid trying to map the
   579  		// region over it (which will cause the kernel to put
   580  		// the region somewhere else, likely at a high
   581  		// address).
   582  		procBrk := sbrk0()
   583  
   584  		// If we ask for the end of the data segment but the
   585  		// operating system requires a little more space
   586  		// before we can start allocating, it will give out a
   587  		// slightly higher pointer. Except QEMU, which is
   588  		// buggy, as usual: it won't adjust the pointer
   589  		// upward. So adjust it upward a little bit ourselves:
   590  		// 1/4 MB to get away from the running binary image.
   591  		p := firstmoduledata.end
   592  		if p < procBrk {
   593  			p = procBrk
   594  		}
   595  		if mheap_.heapArenaAlloc.next <= p && p < mheap_.heapArenaAlloc.end {
   596  			p = mheap_.heapArenaAlloc.end
   597  		}
   598  		p = alignUp(p+(256<<10), heapArenaBytes)
   599  		// Because we're worried about fragmentation on
   600  		// 32-bit, we try to make a large initial reservation.
   601  		arenaSizes := []uintptr{
   602  			512 << 20,
   603  			256 << 20,
   604  			128 << 20,
   605  		}
   606  		for _, arenaSize := range arenaSizes {
   607  			a, size := sysReserveAligned(unsafe.Pointer(p), arenaSize, heapArenaBytes)
   608  			if a != nil {
   609  				mheap_.arena.init(uintptr(a), size, false)
   610  				p = mheap_.arena.end // For hint below
   611  				break
   612  			}
   613  		}
   614  		hint := (*arenaHint)(mheap_.arenaHintAlloc.alloc())
   615  		hint.addr = p
   616  		hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
   617  	}
   618  }
   619  
   620  // sysAlloc allocates heap arena space for at least n bytes. The
   621  // returned pointer is always heapArenaBytes-aligned and backed by
   622  // h.arenas metadata. The returned size is always a multiple of
   623  // heapArenaBytes. sysAlloc returns nil on failure.
   624  // There is no corresponding free function.
   625  //
   626  // sysAlloc returns a memory region in the Reserved state. This region must
   627  // be transitioned to Prepared and then Ready before use.
   628  //
   629  // h must be locked.
   630  func (h *mheap) sysAlloc(n uintptr) (v unsafe.Pointer, size uintptr) {
   631  	assertLockHeld(&h.lock)
   632  
   633  	n = alignUp(n, heapArenaBytes)
   634  
   635  	// First, try the arena pre-reservation.
   636  	v = h.arena.alloc(n, heapArenaBytes, &memstats.heap_sys)
   637  	if v != nil {
   638  		size = n
   639  		goto mapped
   640  	}
   641  
   642  	// Try to grow the heap at a hint address.
   643  	for h.arenaHints != nil {
   644  		hint := h.arenaHints
   645  		p := hint.addr
   646  		if hint.down {
   647  			p -= n
   648  		}
   649  		if p+n < p {
   650  			// We can't use this, so don't ask.
   651  			v = nil
   652  		} else if arenaIndex(p+n-1) >= 1<<arenaBits {
   653  			// Outside addressable heap. Can't use.
   654  			v = nil
   655  		} else {
   656  			v = sysReserve(unsafe.Pointer(p), n)
   657  		}
   658  		if p == uintptr(v) {
   659  			// Success. Update the hint.
   660  			if !hint.down {
   661  				p += n
   662  			}
   663  			hint.addr = p
   664  			size = n
   665  			break
   666  		}
   667  		// Failed. Discard this hint and try the next.
   668  		//
   669  		// TODO: This would be cleaner if sysReserve could be
   670  		// told to only return the requested address. In
   671  		// particular, this is already how Windows behaves, so
   672  		// it would simplify things there.
   673  		if v != nil {
   674  			sysFree(v, n, nil)
   675  		}
   676  		h.arenaHints = hint.next
   677  		h.arenaHintAlloc.free(unsafe.Pointer(hint))
   678  	}
   679  
   680  	if size == 0 {
   681  		if raceenabled {
   682  			// The race detector assumes the heap lives in
   683  			// [0x00c000000000, 0x00e000000000), but we
   684  			// just ran out of hints in this region. Give
   685  			// a nice failure.
   686  			throw("too many address space collisions for -race mode")
   687  		}
   688  
   689  		// All of the hints failed, so we'll take any
   690  		// (sufficiently aligned) address the kernel will give
   691  		// us.
   692  		v, size = sysReserveAligned(nil, n, heapArenaBytes)
   693  		if v == nil {
   694  			return nil, 0
   695  		}
   696  
   697  		// Create new hints for extending this region.
   698  		hint := (*arenaHint)(h.arenaHintAlloc.alloc())
   699  		hint.addr, hint.down = uintptr(v), true
   700  		hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
   701  		hint = (*arenaHint)(h.arenaHintAlloc.alloc())
   702  		hint.addr = uintptr(v) + size
   703  		hint.next, mheap_.arenaHints = mheap_.arenaHints, hint
   704  	}
   705  
   706  	// Check for bad pointers or pointers we can't use.
   707  	{
   708  		var bad string
   709  		p := uintptr(v)
   710  		if p+size < p {
   711  			bad = "region exceeds uintptr range"
   712  		} else if arenaIndex(p) >= 1<<arenaBits {
   713  			bad = "base outside usable address space"
   714  		} else if arenaIndex(p+size-1) >= 1<<arenaBits {
   715  			bad = "end outside usable address space"
   716  		}
   717  		if bad != "" {
   718  			// This should be impossible on most architectures,
   719  			// but it would be really confusing to debug.
   720  			print("runtime: memory allocated by OS [", hex(p), ", ", hex(p+size), ") not in usable address space: ", bad, "\n")
   721  			throw("memory reservation exceeds address space limit")
   722  		}
   723  	}
   724  
   725  	if uintptr(v)&(heapArenaBytes-1) != 0 {
   726  		throw("misrounded allocation in sysAlloc")
   727  	}
   728  
   729  mapped:
   730  	// Create arena metadata.
   731  	for ri := arenaIndex(uintptr(v)); ri <= arenaIndex(uintptr(v)+size-1); ri++ {
   732  		l2 := h.arenas[ri.l1()]
   733  		if l2 == nil {
   734  			// Allocate an L2 arena map.
   735  			l2 = (*[1 << arenaL2Bits]*heapArena)(persistentalloc(unsafe.Sizeof(*l2), sys.PtrSize, nil))
   736  			if l2 == nil {
   737  				throw("out of memory allocating heap arena map")
   738  			}
   739  			atomic.StorepNoWB(unsafe.Pointer(&h.arenas[ri.l1()]), unsafe.Pointer(l2))
   740  		}
   741  
   742  		if l2[ri.l2()] != nil {
   743  			throw("arena already initialized")
   744  		}
   745  		var r *heapArena
   746  		r = (*heapArena)(h.heapArenaAlloc.alloc(unsafe.Sizeof(*r), sys.PtrSize, &memstats.gcMiscSys))
   747  		if r == nil {
   748  			r = (*heapArena)(persistentalloc(unsafe.Sizeof(*r), sys.PtrSize, &memstats.gcMiscSys))
   749  			if r == nil {
   750  				throw("out of memory allocating heap arena metadata")
   751  			}
   752  		}
   753  
   754  		// Add the arena to the arenas list.
   755  		if len(h.allArenas) == cap(h.allArenas) {
   756  			size := 2 * uintptr(cap(h.allArenas)) * sys.PtrSize
   757  			if size == 0 {
   758  				size = physPageSize
   759  			}
   760  			newArray := (*notInHeap)(persistentalloc(size, sys.PtrSize, &memstats.gcMiscSys))
   761  			if newArray == nil {
   762  				throw("out of memory allocating allArenas")
   763  			}
   764  			oldSlice := h.allArenas
   765  			*(*notInHeapSlice)(unsafe.Pointer(&h.allArenas)) = notInHeapSlice{newArray, len(h.allArenas), int(size / sys.PtrSize)}
   766  			copy(h.allArenas, oldSlice)
   767  			// Do not free the old backing array because
   768  			// there may be concurrent readers. Since we
   769  			// double the array each time, this can lead
   770  			// to at most 2x waste.
   771  		}
   772  		h.allArenas = h.allArenas[:len(h.allArenas)+1]
   773  		h.allArenas[len(h.allArenas)-1] = ri
   774  
   775  		// Store atomically just in case an object from the
   776  		// new heap arena becomes visible before the heap lock
   777  		// is released (which shouldn't happen, but there's
   778  		// little downside to this).
   779  		atomic.StorepNoWB(unsafe.Pointer(&l2[ri.l2()]), unsafe.Pointer(r))
   780  	}
   781  
   782  	// Tell the race detector about the new heap memory.
   783  	if raceenabled {
   784  		racemapshadow(v, size)
   785  	}
   786  
   787  	return
   788  }
   789  
   790  // sysReserveAligned is like sysReserve, but the returned pointer is
   791  // aligned to align bytes. It may reserve either n or n+align bytes,
   792  // so it returns the size that was reserved.
   793  func sysReserveAligned(v unsafe.Pointer, size, align uintptr) (unsafe.Pointer, uintptr) {
   794  	// Since the alignment is rather large in uses of this
   795  	// function, we're not likely to get it by chance, so we ask
   796  	// for a larger region and remove the parts we don't need.
   797  	retries := 0
   798  retry:
   799  	p := uintptr(sysReserve(v, size+align))
   800  	switch {
   801  	case p == 0:
   802  		return nil, 0
   803  	case p&(align-1) == 0:
   804  		// We got lucky and got an aligned region, so we can
   805  		// use the whole thing.
   806  		return unsafe.Pointer(p), size + align
   807  	case GOOS == "windows":
   808  		// On Windows we can't release pieces of a
   809  		// reservation, so we release the whole thing and
   810  		// re-reserve the aligned sub-region. This may race,
   811  		// so we may have to try again.
   812  		sysFree(unsafe.Pointer(p), size+align, nil)
   813  		p = alignUp(p, align)
   814  		p2 := sysReserve(unsafe.Pointer(p), size)
   815  		if p != uintptr(p2) {
   816  			// Must have raced. Try again.
   817  			sysFree(p2, size, nil)
   818  			if retries++; retries == 100 {
   819  				throw("failed to allocate aligned heap memory; too many retries")
   820  			}
   821  			goto retry
   822  		}
   823  		// Success.
   824  		return p2, size
   825  	default:
   826  		// Trim off the unaligned parts.
   827  		pAligned := alignUp(p, align)
   828  		sysFree(unsafe.Pointer(p), pAligned-p, nil)
   829  		end := pAligned + size
   830  		endLen := (p + size + align) - end
   831  		if endLen > 0 {
   832  			sysFree(unsafe.Pointer(end), endLen, nil)
   833  		}
   834  		return unsafe.Pointer(pAligned), size
   835  	}
   836  }
   837  
   838  // base address for all 0-byte allocations
   839  var zerobase uintptr
   840  
   841  // nextFreeFast returns the next free object if one is quickly available.
   842  // Otherwise it returns 0.
   843  func nextFreeFast(s *mspan) gclinkptr {
   844  	theBit := sys.Ctz64(s.allocCache) // Is there a free object in the allocCache?
   845  	if theBit < 64 {
   846  		result := s.freeindex + uintptr(theBit)
   847  		if result < s.nelems {
   848  			freeidx := result + 1
   849  			if freeidx%64 == 0 && freeidx != s.nelems {
   850  				return 0
   851  			}
   852  			s.allocCache >>= uint(theBit + 1)
   853  			s.freeindex = freeidx
   854  			s.allocCount++
   855  			return gclinkptr(result*s.elemsize + s.base())
   856  		}
   857  	}
   858  	return 0
   859  }
   860  
   861  // nextFree returns the next free object from the cached span if one is available.
   862  // Otherwise it refills the cache with a span with an available object and
   863  // returns that object along with a flag indicating that this was a heavy
   864  // weight allocation. If it is a heavy weight allocation the caller must
   865  // determine whether a new GC cycle needs to be started or if the GC is active
   866  // whether this goroutine needs to assist the GC.
   867  //
   868  // Must run in a non-preemptible context since otherwise the owner of
   869  // c could change.
   870  func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) {
   871  	s = c.alloc[spc]
   872  	shouldhelpgc = false
   873  	freeIndex := s.nextFreeIndex()
   874  	if freeIndex == s.nelems {
   875  		// The span is full.
   876  		if uintptr(s.allocCount) != s.nelems {
   877  			println("runtime: s.allocCount=", s.allocCount, "s.nelems=", s.nelems)
   878  			throw("s.allocCount != s.nelems && freeIndex == s.nelems")
   879  		}
   880  		c.refill(spc)
   881  		shouldhelpgc = true
   882  		s = c.alloc[spc]
   883  
   884  		freeIndex = s.nextFreeIndex()
   885  	}
   886  
   887  	if freeIndex >= s.nelems {
   888  		throw("freeIndex is not valid")
   889  	}
   890  
   891  	v = gclinkptr(freeIndex*s.elemsize + s.base())
   892  	s.allocCount++
   893  	if uintptr(s.allocCount) > s.nelems {
   894  		println("s.allocCount=", s.allocCount, "s.nelems=", s.nelems)
   895  		throw("s.allocCount > s.nelems")
   896  	}
   897  	return
   898  }
   899  
   900  // Allocate an object of size bytes.
   901  // Small objects are allocated from the per-P cache's free lists.
   902  // Large objects (> 32 kB) are allocated straight from the heap.
   903  func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
   904  	if gcphase == _GCmarktermination {
   905  		throw("mallocgc called with gcphase == _GCmarktermination")
   906  	}
   907  
   908  	if size == 0 {
   909  		return unsafe.Pointer(&zerobase)
   910  	}
   911  
   912  	if debug.malloc {
   913  		if debug.sbrk != 0 {
   914  			align := uintptr(16)
   915  			if typ != nil {
   916  				// TODO(austin): This should be just
   917  				//   align = uintptr(typ.align)
   918  				// but that's only 4 on 32-bit platforms,
   919  				// even if there's a uint64 field in typ (see #599).
   920  				// This causes 64-bit atomic accesses to panic.
   921  				// Hence, we use stricter alignment that matches
   922  				// the normal allocator better.
   923  				if size&7 == 0 {
   924  					align = 8
   925  				} else if size&3 == 0 {
   926  					align = 4
   927  				} else if size&1 == 0 {
   928  					align = 2
   929  				} else {
   930  					align = 1
   931  				}
   932  			}
   933  			return persistentalloc(size, align, &memstats.other_sys)
   934  		}
   935  
   936  		if inittrace.active && inittrace.id == getg().goid {
   937  			// Init functions are executed sequentially in a single goroutine.
   938  			inittrace.allocs += 1
   939  		}
   940  	}
   941  
   942  	// assistG is the G to charge for this allocation, or nil if
   943  	// GC is not currently active.
   944  	var assistG *g
   945  	if gcBlackenEnabled != 0 {
   946  		// Charge the current user G for this allocation.
   947  		assistG = getg()
   948  		if assistG.m.curg != nil {
   949  			assistG = assistG.m.curg
   950  		}
   951  		// Charge the allocation against the G. We'll account
   952  		// for internal fragmentation at the end of mallocgc.
   953  		assistG.gcAssistBytes -= int64(size)
   954  
   955  		if assistG.gcAssistBytes < 0 {
   956  			// This G is in debt. Assist the GC to correct
   957  			// this before allocating. This must happen
   958  			// before disabling preemption.
   959  			gcAssistAlloc(assistG)
   960  		}
   961  	}
   962  
   963  	// Set mp.mallocing to keep from being preempted by GC.
   964  	mp := acquirem()
   965  	if mp.mallocing != 0 {
   966  		throw("malloc deadlock")
   967  	}
   968  	if mp.gsignal == getg() {
   969  		throw("malloc during signal")
   970  	}
   971  	mp.mallocing = 1
   972  
   973  	shouldhelpgc := false
   974  	dataSize := size
   975  	c := getMCache()
   976  	if c == nil {
   977  		throw("mallocgc called without a P or outside bootstrapping")
   978  	}
   979  	var span *mspan
   980  	var x unsafe.Pointer
   981  	noscan := typ == nil || typ.ptrdata == 0
   982  	// In some cases block zeroing can profitably (for latency reduction purposes)
   983  	// be delayed till preemption is possible; isZeroed tracks that state.
   984  	isZeroed := true
   985  	if size <= maxSmallSize {
   986  		if noscan && size < maxTinySize {
   987  			// Tiny allocator.
   988  			//
   989  			// Tiny allocator combines several tiny allocation requests
   990  			// into a single memory block. The resulting memory block
   991  			// is freed when all subobjects are unreachable. The subobjects
   992  			// must be noscan (don't have pointers), this ensures that
   993  			// the amount of potentially wasted memory is bounded.
   994  			//
   995  			// Size of the memory block used for combining (maxTinySize) is tunable.
   996  			// Current setting is 16 bytes, which relates to 2x worst case memory
   997  			// wastage (when all but one subobjects are unreachable).
   998  			// 8 bytes would result in no wastage at all, but provides less
   999  			// opportunities for combining.
  1000  			// 32 bytes provides more opportunities for combining,
  1001  			// but can lead to 4x worst case wastage.
  1002  			// The best case winning is 8x regardless of block size.
  1003  			//
  1004  			// Objects obtained from tiny allocator must not be freed explicitly.
  1005  			// So when an object will be freed explicitly, we ensure that
  1006  			// its size >= maxTinySize.
  1007  			//
  1008  			// SetFinalizer has a special case for objects potentially coming
  1009  			// from tiny allocator, it such case it allows to set finalizers
  1010  			// for an inner byte of a memory block.
  1011  			//
  1012  			// The main targets of tiny allocator are small strings and
  1013  			// standalone escaping variables. On a json benchmark
  1014  			// the allocator reduces number of allocations by ~12% and
  1015  			// reduces heap size by ~20%.
  1016  			off := c.tinyoffset
  1017  			// Align tiny pointer for required (conservative) alignment.
  1018  			if size&7 == 0 {
  1019  				off = alignUp(off, 8)
  1020  			} else if sys.PtrSize == 4 && size == 12 {
  1021  				// Conservatively align 12-byte objects to 8 bytes on 32-bit
  1022  				// systems so that objects whose first field is a 64-bit
  1023  				// value is aligned to 8 bytes and does not cause a fault on
  1024  				// atomic access. See issue 37262.
  1025  				// TODO(mknyszek): Remove this workaround if/when issue 36606
  1026  				// is resolved.
  1027  				off = alignUp(off, 8)
  1028  			} else if size&3 == 0 {
  1029  				off = alignUp(off, 4)
  1030  			} else if size&1 == 0 {
  1031  				off = alignUp(off, 2)
  1032  			}
  1033  			if off+size <= maxTinySize && c.tiny != 0 {
  1034  				// The object fits into existing tiny block.
  1035  				x = unsafe.Pointer(c.tiny + off)
  1036  				c.tinyoffset = off + size
  1037  				c.tinyAllocs++
  1038  				mp.mallocing = 0
  1039  				releasem(mp)
  1040  				return x
  1041  			}
  1042  			// Allocate a new maxTinySize block.
  1043  			span = c.alloc[tinySpanClass]
  1044  			v := nextFreeFast(span)
  1045  			if v == 0 {
  1046  				v, span, shouldhelpgc = c.nextFree(tinySpanClass)
  1047  			}
  1048  			x = unsafe.Pointer(v)
  1049  			(*[2]uint64)(x)[0] = 0
  1050  			(*[2]uint64)(x)[1] = 0
  1051  			// See if we need to replace the existing tiny block with the new one
  1052  			// based on amount of remaining free space.
  1053  			if !raceenabled && (size < c.tinyoffset || c.tiny == 0) {
  1054  				// Note: disabled when race detector is on, see comment near end of this function.
  1055  				c.tiny = uintptr(x)
  1056  				c.tinyoffset = size
  1057  			}
  1058  			size = maxTinySize
  1059  		} else {
  1060  			var sizeclass uint8
  1061  			if size <= smallSizeMax-8 {
  1062  				sizeclass = size_to_class8[divRoundUp(size, smallSizeDiv)]
  1063  			} else {
  1064  				sizeclass = size_to_class128[divRoundUp(size-smallSizeMax, largeSizeDiv)]
  1065  			}
  1066  			size = uintptr(class_to_size[sizeclass])
  1067  			spc := makeSpanClass(sizeclass, noscan)
  1068  			span = c.alloc[spc]
  1069  			v := nextFreeFast(span)
  1070  			if v == 0 {
  1071  				v, span, shouldhelpgc = c.nextFree(spc)
  1072  			}
  1073  			x = unsafe.Pointer(v)
  1074  			if needzero && span.needzero != 0 {
  1075  				memclrNoHeapPointers(unsafe.Pointer(v), size)
  1076  			}
  1077  		}
  1078  	} else {
  1079  		shouldhelpgc = true
  1080  		// For large allocations, keep track of zeroed state so that
  1081  		// bulk zeroing can be happen later in a preemptible context.
  1082  		span, isZeroed = c.allocLarge(size, needzero && !noscan, noscan)
  1083  		span.freeindex = 1
  1084  		span.allocCount = 1
  1085  		x = unsafe.Pointer(span.base())
  1086  		size = span.elemsize
  1087  	}
  1088  
  1089  	var scanSize uintptr
  1090  	if !noscan {
  1091  		// If allocating a defer+arg block, now that we've picked a malloc size
  1092  		// large enough to hold everything, cut the "asked for" size down to
  1093  		// just the defer header, so that the GC bitmap will record the arg block
  1094  		// as containing nothing at all (as if it were unused space at the end of
  1095  		// a malloc block caused by size rounding).
  1096  		// The defer arg areas are scanned as part of scanstack.
  1097  		if typ == deferType {
  1098  			dataSize = unsafe.Sizeof(_defer{})
  1099  		}
  1100  		heapBitsSetType(uintptr(x), size, dataSize, typ)
  1101  		if dataSize > typ.size {
  1102  			// Array allocation. If there are any
  1103  			// pointers, GC has to scan to the last
  1104  			// element.
  1105  			if typ.ptrdata != 0 {
  1106  				scanSize = dataSize - typ.size + typ.ptrdata
  1107  			}
  1108  		} else {
  1109  			scanSize = typ.ptrdata
  1110  		}
  1111  		c.scanAlloc += scanSize
  1112  	}
  1113  
  1114  	// Ensure that the stores above that initialize x to
  1115  	// type-safe memory and set the heap bits occur before
  1116  	// the caller can make x observable to the garbage
  1117  	// collector. Otherwise, on weakly ordered machines,
  1118  	// the garbage collector could follow a pointer to x,
  1119  	// but see uninitialized memory or stale heap bits.
  1120  	publicationBarrier()
  1121  
  1122  	// Allocate black during GC.
  1123  	// All slots hold nil so no scanning is needed.
  1124  	// This may be racing with GC so do it atomically if there can be
  1125  	// a race marking the bit.
  1126  	if gcphase != _GCoff {
  1127  		gcmarknewobject(span, uintptr(x), size, scanSize)
  1128  	}
  1129  
  1130  	if raceenabled {
  1131  		racemalloc(x, size)
  1132  	}
  1133  
  1134  	if msanenabled {
  1135  		msanmalloc(x, size)
  1136  	}
  1137  
  1138  	if rate := MemProfileRate; rate > 0 {
  1139  		// Note cache c only valid while m acquired; see #47302
  1140  		if rate != 1 && size < c.nextSample {
  1141  			c.nextSample -= size
  1142  		} else {
  1143  			profilealloc(mp, x, size)
  1144  		}
  1145  	}
  1146  	mp.mallocing = 0
  1147  	releasem(mp)
  1148  
  1149  	// Pointerfree data can be zeroed late in a context where preemption can occur.
  1150  	// x will keep the memory alive.
  1151  	if !isZeroed && needzero {
  1152  		memclrNoHeapPointersChunked(size, x) // This is a possible preemption point: see #47302
  1153  	}
  1154  
  1155  	if debug.malloc {
  1156  		if debug.allocfreetrace != 0 {
  1157  			tracealloc(x, size, typ)
  1158  		}
  1159  
  1160  		if inittrace.active && inittrace.id == getg().goid {
  1161  			// Init functions are executed sequentially in a single goroutine.
  1162  			inittrace.bytes += uint64(size)
  1163  		}
  1164  	}
  1165  
  1166  	if assistG != nil {
  1167  		// Account for internal fragmentation in the assist
  1168  		// debt now that we know it.
  1169  		assistG.gcAssistBytes -= int64(size - dataSize)
  1170  	}
  1171  
  1172  	if shouldhelpgc {
  1173  		if t := (gcTrigger{kind: gcTriggerHeap}); t.test() {
  1174  			gcStart(t)
  1175  		}
  1176  	}
  1177  
  1178  	if raceenabled && noscan && dataSize < maxTinySize {
  1179  		// Pad tinysize allocations so they are aligned with the end
  1180  		// of the tinyalloc region. This ensures that any arithmetic
  1181  		// that goes off the top end of the object will be detectable
  1182  		// by checkptr (issue 38872).
  1183  		// Note that we disable tinyalloc when raceenabled for this to work.
  1184  		// TODO: This padding is only performed when the race detector
  1185  		// is enabled. It would be nice to enable it if any package
  1186  		// was compiled with checkptr, but there's no easy way to
  1187  		// detect that (especially at compile time).
  1188  		// TODO: enable this padding for all allocations, not just
  1189  		// tinyalloc ones. It's tricky because of pointer maps.
  1190  		// Maybe just all noscan objects?
  1191  		x = add(x, size-dataSize)
  1192  	}
  1193  
  1194  	return x
  1195  }
  1196  
  1197  // memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers
  1198  // on chunks of the buffer to be zeroed, with opportunities for preemption
  1199  // along the way.  memclrNoHeapPointers contains no safepoints and also
  1200  // cannot be preemptively scheduled, so this provides a still-efficient
  1201  // block copy that can also be preempted on a reasonable granularity.
  1202  //
  1203  // Use this with care; if the data being cleared is tagged to contain
  1204  // pointers, this allows the GC to run before it is all cleared.
  1205  func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) {
  1206  	v := uintptr(x)
  1207  	// got this from benchmarking. 128k is too small, 512k is too large.
  1208  	const chunkBytes = 256 * 1024
  1209  	vsize := v + size
  1210  	for voff := v; voff < vsize; voff = voff + chunkBytes {
  1211  		if getg().preempt {
  1212  			// may hold locks, e.g., profiling
  1213  			goschedguarded()
  1214  		}
  1215  		// clear min(avail, lump) bytes
  1216  		n := vsize - voff
  1217  		if n > chunkBytes {
  1218  			n = chunkBytes
  1219  		}
  1220  		memclrNoHeapPointers(unsafe.Pointer(voff), n)
  1221  	}
  1222  }
  1223  
  1224  // implementation of new builtin
  1225  // compiler (both frontend and SSA backend) knows the signature
  1226  // of this function
  1227  func newobject(typ *_type) unsafe.Pointer {
  1228  	return mallocgc(typ.size, typ, true)
  1229  }
  1230  
  1231  //go:linkname reflect_unsafe_New reflect.unsafe_New
  1232  func reflect_unsafe_New(typ *_type) unsafe.Pointer {
  1233  	return mallocgc(typ.size, typ, true)
  1234  }
  1235  
  1236  //go:linkname reflectlite_unsafe_New internal/reflectlite.unsafe_New
  1237  func reflectlite_unsafe_New(typ *_type) unsafe.Pointer {
  1238  	return mallocgc(typ.size, typ, true)
  1239  }
  1240  
  1241  // newarray allocates an array of n elements of type typ.
  1242  func newarray(typ *_type, n int) unsafe.Pointer {
  1243  	if n == 1 {
  1244  		return mallocgc(typ.size, typ, true)
  1245  	}
  1246  	mem, overflow := math.MulUintptr(typ.size, uintptr(n))
  1247  	if overflow || mem > maxAlloc || n < 0 {
  1248  		panic(plainError("runtime: allocation size out of range"))
  1249  	}
  1250  	return mallocgc(mem, typ, true)
  1251  }
  1252  
  1253  //go:linkname reflect_unsafe_NewArray reflect.unsafe_NewArray
  1254  func reflect_unsafe_NewArray(typ *_type, n int) unsafe.Pointer {
  1255  	return newarray(typ, n)
  1256  }
  1257  
  1258  func profilealloc(mp *m, x unsafe.Pointer, size uintptr) {
  1259  	c := getMCache()
  1260  	if c == nil {
  1261  		throw("profilealloc called without a P or outside bootstrapping")
  1262  	}
  1263  	c.nextSample = nextSample()
  1264  	mProf_Malloc(x, size)
  1265  }
  1266  
  1267  // nextSample returns the next sampling point for heap profiling. The goal is
  1268  // to sample allocations on average every MemProfileRate bytes, but with a
  1269  // completely random distribution over the allocation timeline; this
  1270  // corresponds to a Poisson process with parameter MemProfileRate. In Poisson
  1271  // processes, the distance between two samples follows the exponential
  1272  // distribution (exp(MemProfileRate)), so the best return value is a random
  1273  // number taken from an exponential distribution whose mean is MemProfileRate.
  1274  func nextSample() uintptr {
  1275  	if MemProfileRate == 1 {
  1276  		// Callers assign our return value to
  1277  		// mcache.next_sample, but next_sample is not used
  1278  		// when the rate is 1. So avoid the math below and
  1279  		// just return something.
  1280  		return 0
  1281  	}
  1282  	if GOOS == "plan9" {
  1283  		// Plan 9 doesn't support floating point in note handler.
  1284  		if g := getg(); g == g.m.gsignal {
  1285  			return nextSampleNoFP()
  1286  		}
  1287  	}
  1288  
  1289  	return uintptr(fastexprand(MemProfileRate))
  1290  }
  1291  
  1292  // fastexprand returns a random number from an exponential distribution with
  1293  // the specified mean.
  1294  func fastexprand(mean int) int32 {
  1295  	// Avoid overflow. Maximum possible step is
  1296  	// -ln(1/(1<<randomBitCount)) * mean, approximately 20 * mean.
  1297  	switch {
  1298  	case mean > 0x7000000:
  1299  		mean = 0x7000000
  1300  	case mean == 0:
  1301  		return 0
  1302  	}
  1303  
  1304  	// Take a random sample of the exponential distribution exp(-mean*x).
  1305  	// The probability distribution function is mean*exp(-mean*x), so the CDF is
  1306  	// p = 1 - exp(-mean*x), so
  1307  	// q = 1 - p == exp(-mean*x)
  1308  	// log_e(q) = -mean*x
  1309  	// -log_e(q)/mean = x
  1310  	// x = -log_e(q) * mean
  1311  	// x = log_2(q) * (-log_e(2)) * mean    ; Using log_2 for efficiency
  1312  	const randomBitCount = 26
  1313  	q := fastrand()%(1<<randomBitCount) + 1
  1314  	qlog := fastlog2(float64(q)) - randomBitCount
  1315  	if qlog > 0 {
  1316  		qlog = 0
  1317  	}
  1318  	const minusLog2 = -0.6931471805599453 // -ln(2)
  1319  	return int32(qlog*(minusLog2*float64(mean))) + 1
  1320  }
  1321  
  1322  // nextSampleNoFP is similar to nextSample, but uses older,
  1323  // simpler code to avoid floating point.
  1324  func nextSampleNoFP() uintptr {
  1325  	// Set first allocation sample size.
  1326  	rate := MemProfileRate
  1327  	if rate > 0x3fffffff { // make 2*rate not overflow
  1328  		rate = 0x3fffffff
  1329  	}
  1330  	if rate != 0 {
  1331  		return uintptr(fastrand() % uint32(2*rate))
  1332  	}
  1333  	return 0
  1334  }
  1335  
  1336  type persistentAlloc struct {
  1337  	base *notInHeap
  1338  	off  uintptr
  1339  }
  1340  
  1341  var globalAlloc struct {
  1342  	mutex
  1343  	persistentAlloc
  1344  }
  1345  
  1346  // persistentChunkSize is the number of bytes we allocate when we grow
  1347  // a persistentAlloc.
  1348  const persistentChunkSize = 256 << 10
  1349  
  1350  // persistentChunks is a list of all the persistent chunks we have
  1351  // allocated. The list is maintained through the first word in the
  1352  // persistent chunk. This is updated atomically.
  1353  var persistentChunks *notInHeap
  1354  
  1355  // Wrapper around sysAlloc that can allocate small chunks.
  1356  // There is no associated free operation.
  1357  // Intended for things like function/type/debug-related persistent data.
  1358  // If align is 0, uses default align (currently 8).
  1359  // The returned memory will be zeroed.
  1360  //
  1361  // Consider marking persistentalloc'd types go:notinheap.
  1362  func persistentalloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer {
  1363  	var p *notInHeap
  1364  	systemstack(func() {
  1365  		p = persistentalloc1(size, align, sysStat)
  1366  	})
  1367  	return unsafe.Pointer(p)
  1368  }
  1369  
  1370  // Must run on system stack because stack growth can (re)invoke it.
  1371  // See issue 9174.
  1372  //go:systemstack
  1373  func persistentalloc1(size, align uintptr, sysStat *sysMemStat) *notInHeap {
  1374  	const (
  1375  		maxBlock = 64 << 10 // VM reservation granularity is 64K on windows
  1376  	)
  1377  
  1378  	if size == 0 {
  1379  		throw("persistentalloc: size == 0")
  1380  	}
  1381  	if align != 0 {
  1382  		if align&(align-1) != 0 {
  1383  			throw("persistentalloc: align is not a power of 2")
  1384  		}
  1385  		if align > _PageSize {
  1386  			throw("persistentalloc: align is too large")
  1387  		}
  1388  	} else {
  1389  		align = 8
  1390  	}
  1391  
  1392  	if size >= maxBlock {
  1393  		return (*notInHeap)(sysAlloc(size, sysStat))
  1394  	}
  1395  
  1396  	mp := acquirem()
  1397  	var persistent *persistentAlloc
  1398  	if mp != nil && mp.p != 0 {
  1399  		persistent = &mp.p.ptr().palloc
  1400  	} else {
  1401  		lock(&globalAlloc.mutex)
  1402  		persistent = &globalAlloc.persistentAlloc
  1403  	}
  1404  	persistent.off = alignUp(persistent.off, align)
  1405  	if persistent.off+size > persistentChunkSize || persistent.base == nil {
  1406  		persistent.base = (*notInHeap)(sysAlloc(persistentChunkSize, &memstats.other_sys))
  1407  		if persistent.base == nil {
  1408  			if persistent == &globalAlloc.persistentAlloc {
  1409  				unlock(&globalAlloc.mutex)
  1410  			}
  1411  			throw("runtime: cannot allocate memory")
  1412  		}
  1413  
  1414  		// Add the new chunk to the persistentChunks list.
  1415  		for {
  1416  			chunks := uintptr(unsafe.Pointer(persistentChunks))
  1417  			*(*uintptr)(unsafe.Pointer(persistent.base)) = chunks
  1418  			if atomic.Casuintptr((*uintptr)(unsafe.Pointer(&persistentChunks)), chunks, uintptr(unsafe.Pointer(persistent.base))) {
  1419  				break
  1420  			}
  1421  		}
  1422  		persistent.off = alignUp(sys.PtrSize, align)
  1423  	}
  1424  	p := persistent.base.add(persistent.off)
  1425  	persistent.off += size
  1426  	releasem(mp)
  1427  	if persistent == &globalAlloc.persistentAlloc {
  1428  		unlock(&globalAlloc.mutex)
  1429  	}
  1430  
  1431  	if sysStat != &memstats.other_sys {
  1432  		sysStat.add(int64(size))
  1433  		memstats.other_sys.add(-int64(size))
  1434  	}
  1435  	return p
  1436  }
  1437  
  1438  // inPersistentAlloc reports whether p points to memory allocated by
  1439  // persistentalloc. This must be nosplit because it is called by the
  1440  // cgo checker code, which is called by the write barrier code.
  1441  //go:nosplit
  1442  func inPersistentAlloc(p uintptr) bool {
  1443  	chunk := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&persistentChunks)))
  1444  	for chunk != 0 {
  1445  		if p >= chunk && p < chunk+persistentChunkSize {
  1446  			return true
  1447  		}
  1448  		chunk = *(*uintptr)(unsafe.Pointer(chunk))
  1449  	}
  1450  	return false
  1451  }
  1452  
  1453  // linearAlloc is a simple linear allocator that pre-reserves a region
  1454  // of memory and then optionally maps that region into the Ready state
  1455  // as needed.
  1456  //
  1457  // The caller is responsible for locking.
  1458  type linearAlloc struct {
  1459  	next   uintptr // next free byte
  1460  	mapped uintptr // one byte past end of mapped space
  1461  	end    uintptr // end of reserved space
  1462  
  1463  	mapMemory bool // transition memory from Reserved to Ready if true
  1464  }
  1465  
  1466  func (l *linearAlloc) init(base, size uintptr, mapMemory bool) {
  1467  	if base+size < base {
  1468  		// Chop off the last byte. The runtime isn't prepared
  1469  		// to deal with situations where the bounds could overflow.
  1470  		// Leave that memory reserved, though, so we don't map it
  1471  		// later.
  1472  		size -= 1
  1473  	}
  1474  	l.next, l.mapped = base, base
  1475  	l.end = base + size
  1476  	l.mapMemory = mapMemory
  1477  }
  1478  
  1479  func (l *linearAlloc) alloc(size, align uintptr, sysStat *sysMemStat) unsafe.Pointer {
  1480  	p := alignUp(l.next, align)
  1481  	if p+size > l.end {
  1482  		return nil
  1483  	}
  1484  	l.next = p + size
  1485  	if pEnd := alignUp(l.next-1, physPageSize); pEnd > l.mapped {
  1486  		if l.mapMemory {
  1487  			// Transition from Reserved to Prepared to Ready.
  1488  			sysMap(unsafe.Pointer(l.mapped), pEnd-l.mapped, sysStat)
  1489  			sysUsed(unsafe.Pointer(l.mapped), pEnd-l.mapped)
  1490  		}
  1491  		l.mapped = pEnd
  1492  	}
  1493  	return unsafe.Pointer(p)
  1494  }
  1495  
  1496  // notInHeap is off-heap memory allocated by a lower-level allocator
  1497  // like sysAlloc or persistentAlloc.
  1498  //
  1499  // In general, it's better to use real types marked as go:notinheap,
  1500  // but this serves as a generic type for situations where that isn't
  1501  // possible (like in the allocators).
  1502  //
  1503  // TODO: Use this as the return type of sysAlloc, persistentAlloc, etc?
  1504  //
  1505  //go:notinheap
  1506  type notInHeap struct{}
  1507  
  1508  func (p *notInHeap) add(bytes uintptr) *notInHeap {
  1509  	return (*notInHeap)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + bytes))
  1510  }
  1511  

View as plain text